Official~42 GB32GB+ VRAMfull

ltx-2.3-22b-dev.safetensors

LTX 2.3 Dev

Full BF16 dev model. Flexible and trainable. 42GB — requires 48GB VRAM or sequential offloading on 32GB.

Released 2026-03-04 · Source: Lightricks/LTX-2.3 (HuggingFace)Initial LTX 2.3 release. v1.1 weights live in the same repo as ltx-2.3-22b-distilled-1.1.safetensors.

Download ltx-2.3-22b-dev.safetensors

Direct HuggingFace download. ~42 GB · Free.

Install path: ComfyUI/models/checkpoints/ + ltx-2.3-22b-dev.safetensors

No 32GB GPU? Try ltx-2.3-22b-dev.safetensors online — free generation included

Skip the ~42 GB download and ComfyUI setup. Generate a 6-second video using this exact model in your browser, ~30 seconds.

Try this model online — free →

Technical details

ltx-2.3-22b-dev.safetensors is the full-precision (BF16) base model from Lightricks — the source weights that every FP8, MXFP8, NVFP4, and distilled variant is derived from. 22B parameters at BF16 means ~42 GB on disk. This is the file you train against, not the file you reach for to make a quick clip.

Unlike the distilled checkpoints, the dev model uses the full 30-step diffusion schedule with CFG > 1, accepts negative prompts meaningfully, and produces noticeably different output when you stack multiple LoRAs. The trade-off is roughly 4× slower inference and a much larger VRAM footprint.

On 32 GB cards (RTX 5090, A6000) you can run inference with ComfyUI's Sequential Offloading enabled — the model streams layers to and from GPU memory during the forward pass. This is slow but works. For comfortable inference you want 48 GB+ (A6000 Ada, H100 PCIe).

When to choose ltx-2.3-22b-dev.safetensors

Pick the dev BF16 model when you are training LoRAs or fine-tuning — quantized checkpoints lose the gradient precision you need. Also pick it when you need maximum fidelity for a hero shot and inference time is not the constraint.

For day-to-day generation, ltx-2.3-22b-distilled-1.1_transformer_only_fp8_scaled.safetensors or ltx-2.3-22b-dev-fp8.safetensors are better defaults — they cut VRAM in half and inference time by ~4×, with a quality delta most users cannot see.

For LoRA training specifically, this file is non-negotiable. The official LoRA training scripts (and community trainers like ai-toolkit) expect BF16 source weights to compute meaningful gradients.

Will this run on my GPU?

Minimum: 32GB VRAM.

GPUVRAMVerdict
RTX 3060 12GB12GBInsufficient VRAM
RTX 4060 Ti / 4070 (16GB)16GBInsufficient VRAM
RTX 4070 Ti SUPER / 4080 (16GB)16GBInsufficient VRAM
RTX 3090 (24GB)24GBInsufficient VRAM
RTX 4090 (24GB)24GBInsufficient VRAM
RTX 5090 / A6000 (32GB+)32GBTight fit

Recommendation: Best for LoRA training and fine-tuning. 42GB — enable Sequential Offloading on 32GB cards.

How to use ltx-2.3-22b-dev.safetensors

  1. Download the file from HuggingFace.
  2. Place it in ComfyUI/models/checkpoints/ inside your ComfyUI directory.
  3. Restart ComfyUI (or refresh the model list from the menu).
  4. Load a compatible workflow — see below.

Don't want to run this locally? Try ltx-2.3-22b-dev.safetensors online with a free generation — no GPU, no install, ~30 seconds per clip.

Common issues

Loaded the model but OOM on first inference on a 32 GB card

Activations + text encoder + VAE exceed remaining VRAM after the 42 GB transformer is loaded. Even with offloading enabled, peak memory during attention can spike. Fix: Enable Sequential Offloading in ComfyUI settings (not just Model Offload). Also drop to the FP4 Gemma text encoder (gemma_3_12B_it_fp4_mixed.safetensors). Reduce resolution to 576p as a fallback. For comfortable inference at 768p or 1024p you need 48 GB+ VRAM.

Download is extremely slow over residential connection

42 GB file served via HuggingFace's xet protocol. Browser downloads max out at one TCP stream. Fix: Use `huggingface-cli download Lightricks/LTX-2.3 ltx-2.3-22b-dev.safetensors --local-dir .` for multi-stream parallel chunks. Or use aria2c with -x 8. Plan for the size — at 100 Mbps this is ~1 hour.

Training script crashes on backward pass with 'mixed precision' error

Loading the BF16 weights in FP16 mixed precision causes overflow in attention. Fix: Use BF16 mixed precision, not FP16, in your training config. PyTorch's autocast(dtype=torch.bfloat16) is the standard setting for LTX training.

ComfyUI doesn't see the file after I downloaded it

Make sure the file is in ComfyUI/models/checkpoints/ (not a subfolder). Restart ComfyUI fully — the menu refresh sometimes misses new files. Filename must match exactly: ltx-2.3-22b-dev.safetensors.

CUDA out of memory error when loading the model

ltx-2.3-22b-dev.safetensors needs ~32GB VRAM minimum. If you're hitting OOM: • Enable Sequential Offloading in ComfyUI settings • Lower the resolution (768×512 instead of 1280×704) — both dimensions must be divisible by 32 • Reduce frame count (65 frames instead of 161) — must be 8n+1 • Use a smaller variant — see Related models below.

What CFG and step count should I use?

The dev (full) model supports CFG guidance. Try CFG=3.5 with 20-30 steps for highest quality. For faster iteration use CFG=2.5 with 15 steps.

Free newsletter

Get notified when LTX 2.3 Dev updates

Occasional updates on what's new in LTX 2.3 — new FP8 quants, LoRAs, IC-LoRA releases — with our hands-on verdict on whether they're worth re-downloading. No fixed cadence.

No spam. Sent occasionally when there's real news. Unsubscribe in one click.

Related models